Customizing detectors with custom rules

Let us assume we are configuring an anomaly detection job in order to detect DNS data exfiltration. Our data contain fields "subdomain" and "highest_registered_domain". We can use a detector that looks like high_info_content(subdomain) over highest_registered_domain. If we run such a job, it is possible that we discover a lot of anomalies on frequently used domains that we have reasons to trust. As security analysts, we are not interested in such anomalies. Ideally, we could instruct the detector to skip results for domains that we consider safe. Using a rule with a scope allows us to achieve this.

First, we need to create a list of our safe domains. Those lists are called filters in machine learning. Filters can be shared across anomaly detection jobs.

We create our filter using the put filter API:

PUT _ml/filters/safe_domains
{
  "description": "Our list of safe domains",
  "items": ["safe.com", "trusted.com"]
}

Now, we can create our anomaly detection job specifying a scope that uses the safe_domains filter for the highest_registered_domain field:

PUT _ml/anomaly_detectors/dns_exfiltration_with_rule
{
  "analysis_config" : {
    "bucket_span":"5m",
    "detectors" :[{
      "function":"high_info_content",
      "field_name": "subdomain",
      "over_field_name": "highest_registered_domain",
      "custom_rules": [{
        "actions": ["skip_result"],
        "scope": {
          "highest_registered_domain": {
            "filter_id": "safe_domains",
            "filter_type": "include"
          }
        }
      }]
    }]
  },
  "data_description" : {
    "time_field":"timestamp"
  }
}

As time advances and we see more data and more results, we might encounter new domains that we want to add in the filter. We can do that by using the update filter API:

POST _ml/filters/safe_domains/_update
{
  "add_items": ["another-safe.com"]
}

Note that we can use any of the partition_field_name, over_field_name, or by_field_name fields in the scope.

In the following example we scope multiple fields:

PUT _ml/anomaly_detectors/scoping_multiple_fields
{
  "analysis_config" : {
    "bucket_span":"5m",
    "detectors" :[{
      "function":"count",
      "partition_field_name": "my_partition",
      "over_field_name": "my_over",
      "by_field_name": "my_by",
      "custom_rules": [{
        "actions": ["skip_result"],
        "scope": {
          "my_partition": {
            "filter_id": "filter_1"
          },
          "my_over": {
            "filter_id": "filter_2"
          },
          "my_by": {
            "filter_id": "filter_3"
          }
        }
      }]
    }]
  },
  "data_description" : {
    "time_field":"timestamp"
  }
}

Such a detector will skip results when the values of all 3 scoped fields are included in the referenced filters.

Imagine a detector that looks for anomalies in CPU utilization. Given a machine that is idle for long enough, small movement in CPU could result in anomalous results where the actual value is quite small, for example, 0.02. Given our knowledge about how CPU utilization behaves we might determine that anomalies with such small actual values are not interesting for investigation.

Let us now configure an anomaly detection job with a rule that will skip results where CPU utilization is less than 0.20.

PUT _ml/anomaly_detectors/cpu_with_rule
{
  "analysis_config" : {
    "bucket_span":"5m",
    "detectors" :[{
      "function":"high_mean",
      "field_name": "cpu_utilization",
      "custom_rules": [{
        "actions": ["skip_result"],
        "conditions": [
          {
            "applies_to": "actual",
            "operator": "lt",
            "value": 0.20
          }
        ]
      }]
    }]
  },
  "data_description" : {
    "time_field":"timestamp"
  }
}

When there are multiple conditions they are combined with a logical and. This is useful when we want the rule to apply to a range. We simply create a rule with two conditions, one for each end of the desired range.

Here is an example where a count detector will skip results when the count is greater than 30 and less than 50:

PUT _ml/anomaly_detectors/rule_with_range
{
  "analysis_config" : {
    "bucket_span":"5m",
    "detectors" :[{
      "function":"count",
      "custom_rules": [{
        "actions": ["skip_result"],
        "conditions": [
          {
            "applies_to": "actual",
            "operator": "gt",
            "value": 30
          },
          {
            "applies_to": "actual",
            "operator": "lt",
            "value": 50
          }
        ]
      }]
    }]
  },
  "data_description" : {
    "time_field":"timestamp"
  }
}

Custom rules only affect results created after the rules were applied. Let us imagine that we have configured an anomaly detection job and it has been running for some time. After observing its results we decide that we can employ rules in order to get rid of some uninteresting results. We can use the update anomaly detection job API to do so. However, the rule we added will only be in effect for any results created from the moment we added the rule onwards. Past results will remain unaffected.

It might appear like using rules is just another way of filtering the data that feeds into an anomaly detection job. For example, a rule that skips results when the partition field value is in a filter sounds equivalent to having a query that filters out such documents. But it is not. There is a fundamental difference. When the data is filtered before reaching a job it is as if they never existed for the job. With rules, the data still reaches the job and affects its behavior (depending on the rule actions).

For example, a rule with the skip_result action means all data will still be modeled. On the other hand, a rule with the skip_model_update action means results will still be created even though the model will not be updated by data matched by a rule.